Working Definition, History, and Intuition of World Models

A friend throws a tennis ball from three meters away. Before you have even registered the ball's spin, your hand has already moved to the right position. This is neither reflex nor luck: your brain simulated the ball's flight path in that fraction of a second, predicted where it would land, and directed your muscles to act in advance.

A world model is the computational version of this mechanism: a sketch of "how the world works" maintained internally by a brain or an AI, allowing an agent to mentally rehearse outcomes before acting in reality.

After this lecture, you should be able to identify what a system predicts and how its predictions are consumed, distinguish rendering, simulation, and planning, and use a compact capability ladder to compare different systems.

The lecture follows the order in which its questions arise:

• Working definition: distinguish renderers, simulators, and planners, then describe a world model through state estimation, transition, emission, and prediction-query interfaces
• Origins of the idea: move from Craik (1943) through predictive coding and the internal model principle to see why acting after internal rehearsal is not a new idea
• Value and current context: ask what world models solve and why video generation, embodied intelligence, and autonomous driving are converging now
• Taxonomy and roadmap: use the four eras and the L1-L5 capability ladder to organize systems already encountered, then enter the project's build path